Purple L1 Milestone Review Panel - MPI

Jones, T

Lawrence Livermore National Laboratory

关键词: Lawrence Livermore National Laboratory;    Tuning;    Documentation;    Efficiency;    Targets;   

DOI  :  10.2172/1036850
RP-ID  :  UCRL-TR-226719
RP-ID  :  W-7405-ENG-48
RP-ID  :  1036850

美国|英语

来源: UNT Digital Library

【 摘 要 】

The MPI deliverables for the Purple system were designed to ensure that applications which depend on MPI benefit from a robust, functionally complete, and high performance MPI. We specifically targeted three categories of MPI validation: robustness, functionally complete, and high performance. These three categories were intended to address the following needs: (1) Robustness - It doesn't matter how fast you arrive at an answer if the answer is wrong. Since any new flagship machine for the DOE complex will have pushed the envelope for scale, tests were designed to investigate behavior at scale. (2) Functionally complete - MPI functionality concerns usually deal more with coverage than concerns over correctness (no doubt a result of the maturity of the specification). We validated the desired interfaces are present and their operation proceeds as expected. (3) High performance - For a software stack to be considered 'high performance' it must efficiently deliver the capabilities of the underlying hardware and provide levels of performance in keeping with the leading machines of the time. LLNL established separate items for each of the three component areas of robustness, functionally complete, and high performance. Included in functionality was a demonstration of scaling to 8192 tasks, a demonstration of scalable memory usage, acceptable documentation, and full MPI-2 minus dynamic tasking. The robustness element for MPI was addressed separately via full MPI application MTBF in the Synthetic Workload (SWL). In November of 2005, a series of tests were performed on Purple in which all MPI performance and functionality Statement of Work items were passed, save one item. A table outlines the performance measurements. The final performance metric, bi-section bandwidth, was achieved in January 2006 (see Table 2). LLNL and IBM undertook an effort to understand the extent of impact for various levels of shortfall on ASC applications while other efforts continued in parallel to bring up the metric up to the target of 45% efficiency for worse case pairings. By using environment tuning, we were able to achieve 47% efficiency for worse case (see Figure 1). Most pairings actually perform much higher. The Robustness category of MPI was demonstrated by the Synthetic Workload application load Stability Test, or SWL-ST. The test results that were documented and archived will be included in the L1 Milestone completion documentation. All MPI related Statement of Work (SOW) target performance objectives have been met. Both MPI-only and Hybrid-MPI codes have successfully met scaling expectations on Purple (including ale3d, yf3d, and other classified applications).